1. Can you state its job in one sentence?
An agent that "helps with finance" can't be judged. An agent that matches supplier invoices to purchase orders can. If the job is vague, nobody can tell whether it is done well, so nobody can safely stop checking it.
What to look for: one defined job, with the information the agent reads and the output it hands over.
2. Can you read what it did, and why?
Every action should leave a record a person can follow: what the agent was asked, what it read and what it decided. Without that record, a mistake is a mystery, and the only safe response is to switch the agent off.
What to look for: a plain-language record your team can open without calling IT.
3. Has it been checked against how your people do the same work?
"It seems fine" is not evidence. Run the agent while people still review every output, and compare its work with what your team would have done. Write down the errors, the rework and the cases where it got stuck.
What to look for: a period of checked work, with the results recorded before anyone widens the agent's role.
4. What does a mistake cost, and can you undo it?
Drafting a reply that a person then sends carries little risk. Paying a supplier carries a lot. How much a wrong action costs, and how easily it can be reversed, decide how much evidence you need first.
What to look for: autonomy given first where mistakes are cheap and easy to reverse, and last where they are not.
5. Does it know when to stop and ask?
An agent you can trust spots the case it was not built for: the unusual amount, the upset customer, the missing document. It passes that case to a named person instead of guessing.
What to look for: written limits, and a clear route to a person when the agent reaches one.
6. Who is accountable, and can you take autonomy back?
When the agent acts alone, a named person still owns the result. Autonomy should also be easy to narrow again when the evidence changes, for example a new supplier, a new product or a run of errors.
What to look for: an owner for each agent, and a regular review where its autonomy is widened, kept or reduced.
What never earns autonomy
Some decisions stay with a person however strong the evidence: hiring and pay, legal commitments, changes to what you promise a customer, and spending above the limits you set. Write these down first. Everything else works inside them.
Why the questions matter
If you skip them, you risk one of two outcomes. Either nobody trusts the agent, so every output is checked twice, or the agent is trusted too early and keeps that trust until its first expensive mistake. The six questions keep you between the two. Autonomy grows as the record grows, and the people who own the work decide when.
At Futuvara, we build, implement and run AI platforms where AI agents do defined work under rules your business sets.
